A digital computer system generally comprises three basic elements, namely, a memory element, an input/output element and a processor element. The memory element stores information in addressable storage locations. This information includes data and instructions for processing the data. The processor element fetches information from the memory element, interprets the information as either an instruction or data, processes the data in accordance with the instructions, and returns the processed data to the memory element. The input/output element, under control of the processor element, also communicates with the memory element to transfer information, including instructions and the data to be processed, to the memory, and to obtain processed data from the memory.
Most modern computing systems are considered "von Neumann" machines, since they are generally constructed according to a paradigm attributed to John von Neumann. Von Neumann machines are characterized by having a processing element, a global memory which stores all information in the system, and a program counter that identifies the location in the global memory of the instruction being executed. The processing element executes one instruction at a time, that is, the instruction identified by the program counter. When the instruction is executed, the program counter is advanced to identify the location of the next instruction to be processed. (In many modern systems, the program counter is actually advanced before the processor has finished processing the current instruction.)
Von Neumann systems are conceptually uncomplicated to design and program, since they do only one operation at a time. A number of advancements have been made to the original von Neumann paradigm to permit the various parts of the system, most notably the various components of the processor, to operate relatively independently and achieve a significant increase in processing speed. One such advancement is pipelining of the various steps in executing an instruction, including instruction fetch, operation code decode (a typical instruction includes an operation code which identifies the operation to be performed, and in most cases one or more operand specifiers, which identify the location in memory of the operands, or data, to be used in executing the instruction), operand fetch, execution (that is, performing the operation set forth in the operation code on the fetched operands), and storing of processed data, which steps are performed relatively independently by separate hardware in the processor. In a pipelined processor, the processor's instruction fetch hardware may be fetching one instruction while other hardware is decoding the operation code of another instruction, fetching the operands of still another instruction, executing yet another instruction, and storing the processed data of a fifth instruction. Since the five steps are performed sequentially, pipelining does not speed up processing of an individual instruction. However, since the processor begins processing of additional instructions before it has finished processing a current instruction, it can speed up processing of a series of instructions.
A pipelined processor is obviously much more complicated than a simple processor in a von Neumann system, as it requires not only the various circuits to perform each of the operations (in a simple von Neumann processor, many circuits could be used to perform several operations), but also control circuits to coordinate the activities of the various operational circuits. However, the speed-up of the system can be dramatic.
More recently, some processors have been provided with execution hardware which includes multiple functional units each being optimized to perform a certain type of mathematical operation. For example, some processors have separate functional units for performing integer arithmetic and floating point arithmetic, since they are processed very differently. Some processors have separate hardware functional units each of which performs one or only several types of mathematical operations, including addition, multiplication, and division operations, and other operations such as branch control and logical operations, all of which can be operating concurrently. This can be helpful in speeding up certain computations, most particularly those in which several functional units may be used concurrently for performing parts of a single computation.
In a von Neumann processor, including those which incorporate pipelining or multiple functional units (or both, since both may be incorporated into a single processor), a single instruction stream operates on a single data stream. That is, each instruction operates on data to enable one calculation at a time. Such processors have been termed "SISD," for "single-instruction/single-data." If a program requires a segment of a program to be used to operate on a number of diverse elements of data to produce a number of calculations, the program causes the processor to loop through that segment for each calculation. In some cases, in which the program segment is short or there are only a few data elements, the time required to perform such a calculation may not be unduly long.
However, for many types of such programs, SISD processors would require a very long time to perform all of the calculations required. Accordingly, processors have been developed which incorporate a large number of processing elements all of which may operate concurrently on the same instruction stream, but with each processing element processing a separate data stream. These processors have been termed "SIMD" processors, for "single-instruction/multiple-data."
SIMD processors are useful in a number of applications, such as image processing, signal processing, artificial intelligence, database operations, and computer simulation of a number of things, such as electronic circuits and fluid dynamics. In image processing, each processing element may be used to perform processing on a pixel ("picture element") of the image to enhance the overall image. In signal processing, the processors concurrently perform a number of the calculations required to perform such computations as the "Fast Fourier transform" of the data defining the signal. In artificial intelligence, the processors perform searches on extensive rule bases representing the stored knowledge of the particular application. Similarly, in database operations, the processors perform searches on the data in the data-base, and may also perform sorting and other operations. In computer simulation of, for example, electronic circuits, each processor may represent one part of the circuit, and the processor's iterative computations indicate the response of the part to signals from other parts of the circuit. Similarly, in simulating fluid dynamics, which can be useful in a number of applications such as weather predication and airplane design, each processor is associated with one point in space, and the calculations provide information about various factors such as fluid flow, temperature, pressure and so forth.
Typical SIMD systems include a SIMD array, which includes the array of processing elements and a router network, a control processor and an input/output component. The input/output component, under control of the control processor, enables data to be transferred into the array for processing and receives processed data from the array for storage, display, and so forth. The control processor also controls the SIMD array, iteratively broadcasting instructions to the processing elements for execution in parallel. The router network enables the processing elements to communicate the results of a calculation to other processing elements for use in future calculations.
Several routing networks have been used in SIMD arrays and others have been proposed. In one routing network, the processing elements are interconnected in a matrix, or mesh, arrangement. In such an arrangement, each processing element is connected to, and communicates with, four "nearest neighbors" to form rows and columns defining the mesh. This arrangement can be somewhat slow if processing elements need to communicate among themselves at random. However, the arrangement is inexpensive and conceptually simple, and may suffice for some types of processing, most notably image processing. The "Massively Parallel Processor" manufactured by Goodyear Aerospace Corporation is an example of a SIMD array having such a routing network.
In another routing network, processing elements are interconnected in a cube or hypercube arrangement, having a selected number of dimensions, for transferring data, in the form of messages, among the processing elements. The arrangement is a "cube" if it only has three dimensions, and a "hypercube" if it has more than three dimensions. U.S. Pat. No. 4,598,400, entitled Method and Apparatus For Routing Message Packets, issued Jul. 1, 1986 to W. Daniel Hillis, and assigned to the assignee of the present application, describes a system having a hypercube routing network. In the system described in the '400 patent, multiple processing elements are connected to a single routing node, and the routing nodes are interconnected in the hypercube.
Another routing arrangement which has been proposed is a crossbar switch, through which each processing element can communicate directly with any of the other processing elements. The crossbar switch provides the most efficient communications of any of the routing networks proposed. However, a crossbar switch also has the most connections and switching elements, and thus is the most expensive and also the most susceptible to failure due to broken connections and faulty switching elements. Thus, crossbar switch arrangements are rarely used, except when the number of processing elements is fairly small, since the complexity of a crossbar switch increases with the square of the number of processing elements.
Yet another routing arrangement is an butterfly network, in which switching is performed through a number of serially-connected stages. Each stage has two inputs, each connected to the outputs of a prior stage or processing elements, has two outputs which may be connected to the inputs of a subsequent stage or processing elements. The "Butterfly" computer system manufactured by Bolt Beranek & Newman uses such a network. A number of other routing networks, such as a Benes network, have been developed based on the butterfly network.